Two-Stage Detection Algorithm for Kiwifruit Leaf Diseases Based on Deep Learning.

Abstract

The prevention and management of crop diseases play an important role in agricultural production, but there are many types of crop diseases and complex causes, and their prevention and identification add difficulties to the process. The traditional methods of identifying diseases mostly rely on human visual and manual inspection, which requires a certain amount of expert knowledge and experience. There are shortcomings such as strong subjectivity and low accuracy. This paper takes the common diseases of kiwifruit as the research object. Based on deep learning and computer vision models, and given the influence of a complex background in actual scenes on the detection of diseases, as well as the shape and size characteristics of diseases, an innovative method of target detection and semantic segmentation was proposed to identify diseases accurately. The main contributions of this research are as follows: We produced the world's first high-quality dataset on kiwifruit. We used the target detection algorithm YOLOX, we stripped the kiwi leaves from the natural background and removed the influencing factors existing in the complex background. Based on the mainstream semantic segmentation networks UNet and DeepLabv3+, the experimental results showed that the ResNet101 network achieved the most effective results in the identification of kiwi diseases, with an accuracy rate of 96.6%. We used the training method of learning rate decay to further improve the training effect without increasing the training cost. After experimental verification, our two-stage disease detection algorithm had the advantages of high accuracy, strong robustness, and wide detection range, which provided a more efficient solution for solving the problem of precise monitoring of crop growth environment parameters.

Plant phenotyping relevance

キウイ葉の病害状態を対象に、深層学習による物体検出・セマンティックセグメンテーション手法を開発・検証しており、植物病害フェノタイプの取得が中心である。

Code and data availability

The paper's kiwifruit leaf disease image dataset (1000 detection images, 2000 segmented images) is paper-specific but explicitly not public; the authors state it is available only upon request. No public code, models, or repository URLs are provided.

No evidence-backed public reproduction asset is currently recorded.